Data Analysis Prompts

I am providing a data export from [source, e.g., CRM, POS system, Google Analytics] and I want to understand its content and potential insights. Business context: [activity and general question, e.g., I want to understand why sales are declining].

Data: [paste data or first few rows with headers]

Proceed in order: describe the structure (columns, types, period covered, granularity), evaluate quality (missing values, likely duplicates, inconsistencies, empty or constant columns (list them precisely)), calculate basic statistics for key columns, then propose 5 business questions this data can answer and 3 questions it CANNOT answer despite appearances (stating the missing data point).

End with the most promising analysis to perform first and the suggested method. Do not draw business conclusions at this stage.

My metric [metric, e.g., conversion rate, average order value, churn] went from [previous value] to [current value] between [period 1] and [period 2]. Available history: [values from previous periods]. Events during this period: [what changed: campaign, pricing, season, incident, redesign].

Before any interpretation, help me verify: is the variation significant given the usual fluctuations of the series (calculate the magnitude of normal variations from history)? Can seasonality explain it (compare to the same period in previous years or cycles)? Has the measurement scope changed (underlying volume, definition, tracking tool)?

Then, list explanatory hypotheses ranked by plausibility, with the specific test or data point that would confirm each. Forbidden: concluding on a cause based solely on temporal coincidence. End with what you would honestly say in one sentence if asked: "So, is this a problem?"

Help me define a dashboard for [scope, e.g., sales team, online store, customer service]. Business objective: [objective]. Decisions this dashboard must inform: [recurring decisions, e.g., where to focus sales effort, when to hire, which products to promote]. Available data: [sources and tools].

For each listed decision, propose the indicator(s) that truly inform it, distinguishing between lagging indicators (results) and leading indicators (actions that can be taken this week). Maximum 8 indicators total: if you exclude any, explain why.

For each selected indicator: unambiguous definition (what exactly is counted, over what period), formula, source, reporting frequency, and the classic interpretation trap for this metric (mix effect, average hiding a distribution, indicator susceptible to manipulation by the team being measured).

Transform my analysis into a presentation for [audience, e.g., board of directors, manager, client]. Initial analysis question: [question]. What I found: [results, figures, findings]. Analysis limitations: [limitations: partial data, short period, possible bias]. What I recommend: [recommendation if any].

Structure the report in this order: the answer to the question in one sentence (main message first, never method first), 2 or 3 supporting findings each with ITS specific figure (not five figures per finding), the recommendation with its cost and expected return, and limitations stated frankly but without undermining the work (a limitation = a usage precaution, not an excuse).

Also prepare: the 3 likely "gotcha" questions from this audience with answers, and the "30-second elevator pitch" version of the message. Format: [desired output format, e.g., slide outline, email, one-page memo].

I want to test [envisaged change, e.g., new price, new landing page, new call script, send time] before rolling it out. Current situation: [current metric and its level]. Available volume: [traffic, calls, sends per week]. Expected outcome: [expected improvement].

Build the protocol: the hypothesis formulated in a falsifiable way ("if X then Y increases by at least Z"), the success criteria and decision threshold set NOW (before seeing results), the test/control split and how to assign it properly, the estimated duration needed based on my volume to distinguish a real effect from noise (explain the calculation simply), and confounding factors to neutralize (seasonality, day of the week, customer mix).

Add the results interpretation grid: in which case I roll out, in which case I stop, in which case I extend, and the conclusions that will be FORBIDDEN to draw even if the numbers are tempting (subgroups discovered after the fact, stopping the test at the first favorable peak).

You are a data analyst. I need to deduplicate a customer file exported from [tool, e.g., CRM, accounting software] before a campaign. Available columns: [list columns, e.g., first name, last name, email, phone, address, creation date].

File extract: [paste 20 to 30 representative rows].

Propose a multi-pass deduplication strategy: first exact duplicates (same values on which columns?), then probable duplicates (spelling variants, same email but different name, differently formatted phone numbers). For each pass, specify the matching key, necessary pre-normalization (lowercase, spaces, accents, country codes), and the risk of false positives.

Constraints: never automatically merge two records based solely on first/last name, always indicate which record to keep (newest? most complete?) and why. Output format: step-by-step action plan, then concrete Google Sheets formulas or Excel steps to apply it, and an estimate of the number of duplicates detected per pass on the provided extract.

I have an export [source, e.g., POS software, online form] where the date column mixes several formats: [real examples, e.g., 03/04/2026, 2026-04-03, April 3 2026, 04-03-26]. I am working in [Excel or Google Sheets].

Help me convert everything to a single format [target format, e.g., YYYY-MM-DD] without corrupting data. Start by identifying all formats present in my sample and identify AMBIGUOUS values (03/04: March 4th or April 3rd?): list them explicitly instead of deciding arbitrarily, and propose a disambiguation criterion (source tool language, value range greater than 12, consistency with another column).

Then provide the formulas or concrete steps for conversion, including a control column (a formula that flags rows where conversion failed or produced a date outside the plausible period [expected period]). End with a list of checks to perform before replacing the original column.

You are a rigorous data analyst. My file [description, e.g., export of annual orders from ERP] contains missing values, and I must decide how to handle them before analysis. Affected columns and approximate null rate: [e.g., zip code 12%, acquisition channel 40%, amount 2%].

Planned analysis: [what I want to calculate, e.g., Revenue by region and channel].

For each column, help me answer in order: is the void informative (non-applicable field, optional entry, collection bug) or purely random? Which option is least risky for MY analysis: excluding rows, imputing a value (which one and why), creating an "Unknown" category, or abandoning that specific analysis axis? What bias does each option introduce?

Constraints: never impute by the mean without justification, always quantify the impact (how many rows lost, what weight in the total). Format: column-by-column table with recommended decision, justification, and note to include in the final report.

I have a free-text column to standardize: [field type, e.g., city names, company names, product categories] in [tool, e.g., Google Sheets, Excel]. Real value examples: [paste 20 to 40 values, including variants: case, abbreviations, typos, spaces].

Objective: obtain a clean column with one canonical value per entity for correct aggregation.

Proceed as follows: identify groups of variants referring to the same entity in my sample, propose the canonical value for each group, then list doubtful cases to be validated manually rather than deciding for me. Then provide the implementation method: mapping table (variant to canonical) plus join formula, upstream normalization rules (case, accents, multiple spaces, punctuation), and a control formula that isolates values left without a match.

Constraint: the mapping table must be a separate tab, reusable for future exports, not manual one-off replacements. Format: mapping table, then ready-to-paste formulas.

I need to import a file of [nature, e.g., trade show contacts, form leads] into [CRM or target tool]. My file columns: [list columns]. Fields expected by the target tool: [list fields and known constraints, e.g., email required, phone in international format].

Sample: [paste 10 to 20 rows].

Prepare the compliance plan: source column to target field mapping (identifying orphaned columns on both sides), checks to pass before import (syntactically invalid emails, incomplete phones, empty required fields, duplicates with explicit criteria), necessary transformations (first/last name separation, phone formatting, country standardization).

Constraints: every rejected row must end up in a "rejects" tab with the reason, never silently deleted; no invented fields to fill a required one. Output format: mapping table, ordered list of checks with the formula or method for each, and final checklist before launching the import (backup, 10-row test import, verification in the tool).

You are a senior SQL developer working with a business stakeholder. My business question: [question, e.g., which customers ordered at least twice this year but nothing in the last three months?]. Database: [engine, e.g., PostgreSQL, MySQL, BigQuery]. Available tables and columns: [describe tables, columns, and relationships, or paste schema].

Before writing the query, reformulate my question in precise terms and list ambiguous points with the assumption you are making for each (exactly what is counted, what are the date boundaries, are canceled orders included, time zone). If essential information is missing from the schema, say so instead of inventing a column.

Then write the query, commented line by line, then propose a simple control query (a total or count) to verify the result is plausible before I rely on it. Format: assumptions, commented query, control query, result limitations.

I inherited a SQL query used in [context, e.g., monthly reporting, automatic export] and no one knows exactly what it does anymore. Here it is: [paste full query].

Known context: [what the report is supposed to show, database engine if known].

Explain it at three levels: a summary sentence (what it calculates in business language), then a block-by-block breakdown (each join, filter, aggregation, and subquery with its concrete effect), then hidden traps: joins that might duplicate or lose rows, WHERE filters that turn a LEFT JOIN into an INNER JOIN, NULL handling, inclusive vs. exclusive date boundaries, implicit data assumptions.

End with: cases where the resulting figure would be WRONG without any warning, and two quick control queries to check these cases on my database. Format: structured explanation, then list of traps ranked by severity.

You are a SQL optimization expert. My query is too slow: [current time] for [approximate volume of affected tables]. Engine: [PostgreSQL, MySQL, SQL Server, BigQuery]. Query: [paste query]. Existing indexes if known: [list or state "unknown"]. Execution plan if available: [paste EXPLAIN output or state "not available"].

Diagnose first: identify likely causes of slowness in the query (functions applied to filtered columns preventing index use, correlated subqueries, unnecessary SELECT *, joins on non-indexed columns, sorting or aggregation on large volumes, lack of date filters).

Then propose optimizations ranked by expected effort/gain ratio: query rewriting (with full rewritten version), indexes to create (with their write cost), and as a last resort, structural changes. Absolute constraint: the result must remain strictly identical; for each rewrite, explain why equivalence is guaranteed. End with the method to properly measure the gain.

I produced a result with this SQL query and I want to check it before sharing it with [recipients, e.g., management, a client]. Query: [paste query]. Result obtained: [figure(s)]. Expected order of magnitude according to the business: [what would be plausible, or "no idea"].

Build my battery of checks: a raw volume query (how many rows before any filter, how many after each main filter, to see where volume drops), a duplicate check on the supposed unique key, a NULL check on join and aggregation columns, a reconciliation with an independent source if I have one [available cross-check source], and a boundary test (does the figure change if I shift the period by one day?).

For each check: provide the ready-to-run query, what a normal result looks like, and what should alert me. End with the decision rule: in which cases I share, correct, or warn that the figure is provisional.

I am working in [Excel or Google Sheets] on a file of [data nature, e.g., sales lines with date, product, store, quantity, amount]. Question I want to answer: [business question, e.g., which products are driving growth and in which stores?].

Available columns: [list columns and content].

Design the appropriate Pivot Table(s): for each, indicate which fields to put in rows, columns, values, and filters, which aggregation to choose (sum, average, count, and why average might be misleading if applicable), and what grouping to apply (dates by month, amount brackets).

Flag the traps in my case: rows that duplicate if granularity isn't what I think, totals mixing units, percentages calculated on the wrong base, empty items hidden by default. Provide click-by-click construction steps in my tool, then two insights the table will allow and one conclusion that would be tempting but false to draw.

In [Excel or Google Sheets], I want to bring information from another tab into my main tab. Main tab: [columns, e.g., A order number, B date, C amount]. Source tab: [columns, e.g., A order number, D customer name, E segment]. Information to bring over: [desired column]. Matching key: [common column].

Give me the exact formula, ready to paste, choosing between VLOOKUP, INDEX/MATCH, and XLOOKUP based on my tool version [specify version if known], and explain your choice in two sentences.

Explicitly handle cases that break these formulas: key missing from source (what to return instead of an error), duplicate keys in source (which row is taken, how to detect it with a count formula), keys that look identical but differ by a space or number/text format (cleaning formula to apply first). Format: main formula, duplicate detection formula, cleaning formula, each with the exact range to adapt.

I am working in [Excel or Google Sheets]. My data: [describe columns with letters, e.g., A date, B sales rep, C region, D amount, E status]. What I want to calculate, in plain English: [e.g., total sales for the North region signed in 2026 by sales rep, excluding 'canceled' status].

Write the exact formula(s) with the correct ranges, choosing between SUMIFS, COUNTIFS, AVERAGEIFS, or a combination. Explain each criterion of the formula in one line.

Cover the subtleties of my case: date criteria (how to write "in 2026" or "between two dates" in a criterion, with quotes and ampersand in the right place), exclusion criteria (not equal to), wildcards for partial text, and absolute references with dollar signs for copying the formula. Add a control formula: the total of my formulas by criterion must equal the grand total; write this check. Format: ready-to-paste formulas, then one line of explanation each.

In Google Sheets (REGEXEXTRACT, REGEXMATCH, REGEXREPLACE functions), I want to extract [target info, e.g., zip code, invoice number, email domain] from a free-text column. Real value examples with their diversity: [paste 10 to 15 representative values, including edge cases].

Expected result for each: [give desired result for 3 or 4 examples].

Write the full formula, then explain the regex piece by piece (each character class, quantifier, and group). Mentally test it against each of my examples and flag those it would miss, rather than promising it works everywhere.

Add: a REGEXMATCH variant to create a true/false control column (is the pattern present?), handling the "no match" case without showing an error, and known limitations of your pattern (uncovered writing variants). Constraint: prioritize a readable and slightly longer regex over an unreadable compact one. Format: ready-to-paste formulas, then explanation.

My [Excel or Google Sheets] formula is not giving the expected result. Exact formula: [paste formula]. Error or symptom: [e.g., #N/A on some rows, #REF!, wrong total, zero everywhere]. Context: [describe columns used and an example of a row where it fails, with real values].

Proceed as a diagnosis, not a blind rewrite: list possible causes for MY symptom ranked from most to least likely (mixed number/text types, invisible spaces, shifted ranges after row insertion, non-fixed reference copied, searched value missing, required sorting not respected, locale separators).

For each cause, provide the test that confirms or eliminates it in less than a minute (ISNUMBER, ISTEXT, LEN formula to see spaces, strict comparison). Once the cause is identified, provide the minimal fix, then the robust version of the formula that avoids recurrence. Constraint: do not propose redo-ing the whole file. Format: diagnostic tree, tests, correction.

I need to present data to [audience, e.g., board, field team] and I want to choose the right chart. Message I want to convey: [message in one sentence, e.g., two products account for most of the growth]. Available data: [describe or paste data: dimensions, measures, number of categories, period].

Recommend the chart type adapted to this MESSAGE (comparison, trend, distribution, relationship) and justify in two sentences. Propose a credible alternative and explain why you rank it second. Explicitly eliminate tempting but poor choices for my data (pie charts with too many categories, dual axes suggesting a relationship, unreadable stacked areas) with the reason.

Then provide concrete specifications: what goes on the x and y axes, category sorting, scale (zero included or not, and why), what to annotate directly on the chart, and a title formulated as the message rather than a description. End with the "what not to do" version: the choice that would distort the message.

Here is a chart I am about to share in [context, e.g., monthly report, client presentation]. Precise description or data of the chart: [describe chart: type, axes, scales, series, title, or paste represented data]. Message it is supposed to carry: [the message].

Critique it in three parts. Honesty: does the scale distort the variation (unflagged truncated axis, dual axis, unannounced log scale), is the chosen period cherry-picked, are sample sizes sufficient for the percentages shown? Readability: what does a reader understand in five seconds, does the title state the message or just the subject, are there elements to remove (colors, grids, decimals, redundant legend)? Alignment: does this chart type serve this message, or would another do better?

Rank your remarks as "blocking before sharing" and "desirable improvement," then precisely describe the corrected version. Constraint: critique the chart, not the data; if the problem is the data, say so separately.

I want to build [target chart, e.g., monthly trend curve by channel, stacked bar chart] in [tool, e.g., Google Sheets, Excel, Looker Studio] but my raw data is not in the right format. Current data: [paste extract with headers, e.g., one row per transaction].

First describe the exact target structure my tool expects for this chart (long vs. wide format, one column per series or a category column, date granularity), with a 5-row example of what my data should look like at the end.

Then provide the simplest transformation path in my tool: aggregation (pivot table or formulas, which ones), handling periods without data (showing them as zero or leaving them out, and the effect on the curve), sorting, and rounding only for display, never in the data.

End with two checks before plotting: the transformed table total equals the raw total, and the number of periods matches the expected range [period]. Format: target structure, steps, checks.

In my company, the indicator [indicator name, e.g., conversion rate, revenue, active customers] gives different figures depending on who calculates it, and I want to write ITS reference definition. Known current uses: [who uses it and how, e.g., marketing calculates it from GA4, sales from CRM]. Available data sources: [tools and tables].

Write the full definition sheet: exact formula (numerator and denominator defined unambiguously: which events or rows count, which are excluded, how edge cases like cancellations, refunds, duplicates, internal tests are handled), reference period and time zone, authoritative source in case of discrepancies between tools, frequency, and freeze date (can a month's figure still move later?).

Then list the likely discrepancies between current uses and this definition, with the expected magnitude of difference, to prepare for the switch. Constraint: every term in the formula must be verifiable by query; no fuzzy terms ("active customer" must be defined by a measurable criterion). Format: structured one-page sheet.

My main indicator is [indicator, e.g., monthly revenue, number of new customers] and I want to decompose it into levers to know where to act. Context: [activity, e.g., e-commerce site, service firm, retail network]. Current level: [current value]. Goal: [target goal].

Build the arithmetic decomposition tree: the indicator expressed as a product or sum of sub-indicators (e.g., traffic multiplied by conversion rate multiplied by average order value), then each sub-indicator decomposed one level further where useful. Verify that each decomposition is mathematically exact, not approximate.

For each leaf of the tree: who in the organization can act on it, the data that allows it to be measured [available sources], and its realistic margin for improvement. Then identify the two levers where a given effort yields the most (sensitivity: what happens to the main indicator if this lever gains 10%?), showing the calculation. Constraint: flag levers that cannibalize each other. Format: indented tree then table of levers.

I want to analyze the retention of my [customers or subscribers] by cohorts. Available data: [describe, e.g., export with customer ID, first order or signup date, subsequent activity or order dates]. Tool: [Excel, Google Sheets, SQL]. Definition of a "retained" customer in my business: [e.g., reordered within the month, has an active subscription].

Guide me in building the cohort table: grouping by month of first order, calculating activity at M+1, M+2, M+3 and beyond, as a percentage of the initial cohort size. Provide the concrete method in my tool (formulas or query).

Then help me READ it: distinguish the tenure effect (all cohorts erode over time) from the vintage effect (are recent cohorts better or worse than old ones at the same age?), identify the month of tenure where retention stabilizes, and flag cohorts that are too young or too small to draw conclusions (with your recommended threshold). Format: construction method, then reading grid with questions to ask.

I want to segment my customer base using the RFM (Recency, Frequency, Monetary) method to target my actions. Data: [available export, e.g., one row per order with customer ID, date, amount, from POS or CRM]. Tool: [Excel, Google Sheets, SQL]. Activity: [type of activity and typical purchase cycle, e.g., fashion e-commerce, purchase every 2 months].

Build the approach: calculation of the three values per customer (with the exact formula or query in my tool), choice of scoring thresholds adapted to MY purchase cycle rather than automatic quintiles (justify each threshold), score attribution, and grouping into 5 to 8 named, meaningful segments (Loyalists, Hibernating, Promising Newcomers, Lost...). 

For each segment: expected size in percentage, one-sentence characterization, priority commercial action, and indicator to measure if it works. Constraints: exclude canceled orders and test customers, handle one-time buyers, and flag that thresholds should be reviewed at [recommended frequency]. Format: step-by-step method then segment table.

I want to forecast [quantity, e.g., monthly sales, bookings, call volume] for the next [horizon, e.g., 6 months]. Available history: [paste series, ideally 24 months or more, or describe what I have]. My activity has known seasonality: [describe, e.g., peak in December, dip in August].

Build a simple and reproducible forecast in [Excel or Google Sheets]: calculate the underlying trend (12-month moving average or regression, explain your choice), monthly seasonality coefficients from history, then the forecast (trend multiplied by coefficient). Show the formulas.

Demand honesty in the result: provide a low/high range built from the errors this method would have made in the past (backtest: apply the method to the last 6 known months and measure the variance), not a decorative range. List what the method CANNOT see (exceptional events, price changes, atypical base year) and the signal that should trigger a mid-course forecast revision. Format: method, formulas, forecast table with range, limitations.

We are in [current month] and I want to project my year-end landing for [quantity, e.g., revenue, expenses, number of cases handled]. Actual to date: [cumulative value]. Annual goal: [goal]. Monthly history from last year and this year: [paste values].

Build three projections and compare them: the naive rule of three (cumulative divided by elapsed months times twelve), the seasonalized projection (the usual weight of remaining months based on history), and the "informed" projection that integrates what I already know about coming months: [known elements, e.g., signed orders, pending quotes, summer closure, hiring].

For each: provide the figure, the implicit assumption it makes, and where it is most likely to be wrong. Recommend which one to use for my case and the gap to goal it implies, in dollars and percentage. End with the two or three actions that could still move the landing and a rough estimate for each. Format: comparative table then 5-line recommendation.

You are a data quality manager. I want to audit the quality of [affected database, e.g., CRM customer base, product catalog, supplier file] before [intended use, e.g., migration, campaign, setting up reporting]. Structure: [list fields]. Representative extract: [paste 30 to 50 rows].

Audit according to these axes, with a quantified finding for each on my extract: completeness (fill rate per field, distinguishing critical fields for my use), validity (invalid formats: emails, phones, zip codes, impossible dates), uniqueness (likely duplicates and their detection criteria), freshness (update dates, share of likely obsolete records), internal consistency (contradicting fields, e.g., zip code and city).

Then produce the remediation plan prioritized by impact on MY intended use: what blocks, what degrades, what can wait, with the correction method (automated or manual) and estimated effort for each batch. Format: audit table by axis, then plan in three batches.

I want to create a data dictionary for [dataset, e.g., ERP order export, customer table in our DB, shared reporting Sheet] so the team stops misinterpreting columns. Columns and extract: [list columns with 5 example rows].

What I know about the context: [data origin, update frequency, who maintains it].

For each column, write the entry: technical name and business label, one-sentence definition without jargon, type and expected format, possible values or plausible range, filling rule (required, calculated, manual entry), upstream source, and especially the "traps" section: what a user might wrongly assume about this column (Gross vs Net, order date vs payment date, historical vs overwritten field). When you cannot deduce information from the extract, write "to be confirmed with [whom]" instead of inventing. 

Format: a table ready to paste into a "Dictionary" tab, plus a list of questions to confirm, sorted by likely owner.

Two of my tools give different figures for the same reality: [Source A, e.g., GA4] reports [Value A] and [Source B, e.g., POS software, Stripe, CRM] reports [Value B] for [metric and period]. Discrepancy: [percentage or amount]. I need to understand the gap and decide which source is authoritative.

Guide the investigation in order: first DEFINITION gaps (exactly what does each tool count: refunds, cancellations, tests, time zones, order date vs. settlement date, bot filters), then SCOPE gaps (channels or locations missing from one side), then COLLECTION gaps (tracking loss, consent, outages). For each hypothesis: provide the concrete test to quantify it, starting with those that usually explain the largest gaps.

Final objective: a bridge table from Value A to Value B, discrepancy item by discrepancy item, that sums up correctly. End with your recommendation: which source to trust for which use, and the residual acceptable gap below which to stop digging.

I am about to share [deliverable, e.g., monthly report, client presentation, email with quarterly results] with [recipients]. Key figures it contains: [list main figures with their labels]. Sources: [where each figure comes from].

Walk me through the final control checklist, point by point, asking me to confirm each: internal consistency (subtotals match totals, percentages and their bases align, figures cited twice are identical everywhere), external consistency (comparison to last month and last year: does any surprising variation have a verified explanation, not just a plausible one?), announced scope (each figure states its period, scope, and currency; definitions haven't changed since the last edition without warning), reproducibility (I know how to recreate each figure: query or source file saved), and rounding (consistent, no illusory precision).

For each point I cannot confirm, indicate the concrete risk and whether it is blocking for publication or not. Format: numbered checklist with checkboxes.

I want to automatically detect abnormal values in [tracked series, e.g., daily sales, hourly orders, daily signups] to react quickly to problems. History: [paste series or describe: granularity, duration, typical values]. Tool: [Google Sheets, Excel, SQL]. Known normal variations: [e.g., strong Saturdays, weak Sundays, seasonality].

Propose a simple and robust setup: a reference value that respects my cycles (comparing a Monday to previous Mondays, not the day before), a deviation measure, and an alert threshold justified by the actual dispersion of my history (show the calculation, explain why this threshold and not an arbitrary percentage), and an anti-false alarm rule (e.g., alert only after two consecutive abnormal points, distinct thresholds for high and low anomalies).

Provide the concrete implementation in my tool, including formulas. End with: the three benign causes that will trigger the alert wrongly (public holiday, promotion, collection outage) and how to handle them, plus the three-step procedure when an alert occurs.

A figure in my reporting seems wrong and I must decide: calculation error, data problem, or surprising reality. The figure: [value and label, e.g., March revenue shows $240,000 vs usual $90,000]. Where it appears: [report, query, or file producing it]. Known production chain: [where data comes from and what transformations it undergoes].

Structure the investigation in layers, from downstream to upstream, with a test at each layer to clear or accuse it: the final calculation (formula, range, report filter), the transformation (aggregation, join duplicating rows, import duplicate, recent definition change), the collection (outage, massive catch-up on a given day, tool change), and finally the business reality (large order, exceptional client: what specific record would prove it?).

Constraint: forbidden to correct anything before identifying the cause; a surprising but true figure is not "corrected." Format: ordered investigation plan, then the three possible outcomes with the associated action for each (correct, document, alert).

I collected open-ended customer responses via [channel, e.g., satisfaction survey open question, Google reviews, cancellation reasons] and I want to derive actionable analysis. Verbatims: [paste verbatims, or a sample if voluminous, with total count]. Question asked to customers: [exact question].

Proceed as follows: build a theme grid from the verbatims themselves (no pre-conceived grid), with a one-line definition per theme to ensure reproducible classification. Classify each verbatim (multi-theme allowed), then quantify: number of mentions per theme, dominant sentiment, and cross-reference with [available variable, e.g., rating given, customer tenure] if provided.

Honesty constraints: quote verbatims textually without rewriting, distinguish frequent themes from rare but serious ones, specify the number of respondents everywhere (no percentages on fewer than 30 responses without flagging it), and do not invent any quotes. Format: one-page summary (themes ranked by frequency, 2 representative quotes per theme, weak signals to watch), then the full classification table.

I ran a survey via [tool, e.g., Google Forms, Typeform] and I need to analyze it. Number of responses: [count]. Questions asked: [list questions with type: single choice, multiple choice, scale, open]. Export: [paste response extract]. Survey objective: [what I am trying to decide or learn].

Walk through the analysis in order: cleaning (duplicate responses, surveys completed in seconds, blatant internal inconsistencies: list what you discard and why), frequency tables for each question (with the appropriate reporting format for each question type), then the 3 to 5 most useful cross-tabs for MY objective (justify the choice, do not cross-tab everything with everything).

Mandatory precautions: remind who responded and how the sample might differ from my target (distribution method, self-selection), flag every cross-tab where a cell falls below 30 respondents, and ban causal conclusions on self-reported data. Format: analysis plan, frequency and cross-tab results on my extract, limitations in a final box.

I implemented [change, e.g., new hours, page redesign, new price, new sales pitch] on [date] and I want to know if it worked, without a control group. Tracked metric: [metric]. Before values: [series before the change, as long as possible]. After values: [series since the change]. Other events during the period: [what changed at the same time: season, campaign, news, weather, competition].

Before comparing averages, force me to verify: is the "before" window representative (long enough, no anomalies), does the observed variation exceed the metric's usual fluctuations (quantify them from history), does seasonality explain the gap (compare to the same prior period if available), and which concurrent event could produce the same effect?

Conclude at three levels: what the data allows us to affirm, what it only suggests, and what it cannot decide. End with the setup that would make the next evaluation more solid (phased rollout, alternation, control group). Forbidden: concluding based solely on temporal coincidence.

Write a one-page memo based on my analysis work, intended for [reader, e.g., director, investment committee, client] who will read it in [available time, e.g., three minutes] and must decide: [expected decision]. Raw material: [paste analysis results, figures, findings, even if messy]. My recommendation if I have one: [recommendation].

Mandatory structure: a title that states the conclusion (not the subject), a first paragraph giving the answer and key figure, then maximum 3 findings each with ONE supporting figure, possible options with their cost and risk in one line each, the stated recommendation, and a "what we don't know yet" line stating limitations without drowning the message.

Style constraints: one page maximum, short sentences, no analyst jargon (no "correlation" or "dataset" without translation), figures rounded to the precision useful for deciding, no figures in the memo not contained in the raw material. If the material is insufficient for a structural element, write [missing data] instead of inventing. 

Help me transform my analysis results into a presentation that sticks, for [audience and occasion, e.g., quarterly review for management, client meeting]. Duration: [duration]. Results to present: [paste key findings and figures]. What I want to get from the audience at the end: [decision, budget, priority shift].

Build the narrative arc: the situation the audience knows and accepts, the tension (the figure or finding that disturbs or surprises), the exploration (why it is this way, 2 or 3 proofs max), the resolution (what I propose), and the call to action. Then break it down into slides: one idea per slide, with each slide title being an affirmative sentence carrying the idea (the sequence of titles alone must tell the whole story: verify this).

For each key figure, propose an embodiment that makes it tangible (comparison to a known quantity for the audience, translation into dollars or customers per week). Constraints: no figures invented for the story, analysis caveats get a slide, not a footnote. Format: slide-by-slide plan with title and content.

Translate my analysis result into a clear explanation for [audience, e.g., sales team in a meeting, franchisees, end client] who have no statistical background. Technical result: [state result with technical terms, e.g., 0.7 correlation between X and Y, confidence interval, A/B test result, scoring model]. What the audience needs to take away to act: [expected action or change].

Produce three levels of explanation: the 30-second verbal version (everyday language, one daily life analogy if it helps, no technical terms), the 5-minute version with a concrete numerical example from my data, and the answer to the 3 questions this audience will likely ask.

Fidelity constraints, non-negotiable: the simplification must not say MORE than the technical result (correlation does not become causation, a trend does not become a certainty, a segment result does not become general); explicitly flag every place you simplified and what the simplification leaves out. Format: the three levels, then the list of assumed simplifications.

I want to document [affected report, e.g., monthly sales dashboard, weekly activity report] so another person can produce, understand, and fix it without me. How it is produced today: [describe mechanics: sources, extractions, files, formulas, manual processing, distribution]. Frequency: [frequency]. Recipients: [who receives it and what they do with it].

Write the documentation in four parts: the production manual (each step in order, with the exact place to click or query to run, including "shameful" manual steps: these are the ones that get lost), the indicator dictionary (definition, formula, source for each), the dependency map (what breaks the report: an export changing format, an expired access, a renamed tab), and the incident protocol (how to verify a contested figure, who to notify, how to flag a wrong report already shared).

Ask me questions if my description has holes, rather than filling with guesses. Format: structured document ready to be stored in [location, e.g., team Drive].

I need to merge two files referring to the same entities but without a reliable common ID: [File 1, e.g., CRM export with company name and email] and [File 2, e.g., accounting export with legal name and invoice number]. File 1 columns: [list]. File 2 columns: [list]. Extracts: [paste 10 rows of each]. Merge objective: [what I want to achieve].

Build the matching strategy in cascades, from most to least certain: first exact keys available after normalization (lowercase email, tax ID, cleaned number), then composite keys (normalized name plus zip code), then fuzzy matching on names (explaining the practicable method in [Excel, Google Sheets] and its limits). For each cascade: how many rows matched on my extracts, and the feared false positive rate.

Constraints: every match keeps a "method" column stating which cascade it passed through; non-matches end up in a manual work list, never silently discarded. Format: cascade plan, formulas, matching rate tracking table.

I cannot examine the entire population and must work on a sample. Population: [description and size, e.g., 12,000 customer files, 3,000 annual invoices, 800 recorded calls]. What I want to estimate or verify: [objective, e.g., data entry error rate, share of compliant files]. Precision needed to decide: [e.g., know if the rate exceeds 5%, an estimate within plus or minus 3 points].

Build the plan: the sample size needed for my target precision (show the reasoning simply, without dropped-in formulas), the drawing method that avoids obvious biases (simple random vs. stratified: if my population has very different subgroups [describe possible strata], recommend and dimension stratification), and the concrete drawing procedure in [Excel or Google Sheets] (reproducible random drawing formula).

List the biases threatening MY case despite the drawing (missing files replaced by easy ones, non-representative period) and the fix for each. End with the honest template sentence to present the final result with its margin of uncertainty.

My conversion rate varies depending on how it's calculated and I want to fix it once and for all. Context: [affected journey, e.g., site visit to purchase, quote to signature, call to appointment]. Measurement sources: [tools, e.g., GA4, CRM, POS]. Currently used calculations and their results: [different versions circulating].

Help me set the rigorous definition: the numerator (which event counts as a conversion, handling cancellations, duplicates, multiple conversions from one person), the denominator (who enters the base: unique visitors or visits, qualified leads or all contacts, on what entry criterion), and the attribution window (time between entry and conversion, and which period to attribute a conversion spanning two months).

For each choice, show the effect on the figure with my data and indicate the most common convention. Explain why my current calculations diverge. End with the selected definition in one publishable sentence, and the two reading errors to prevent in those who see this rate (comparing rates with different scopes, averaging rates without weighting).

I want to calculate and track churn (attrition) for [activity, e.g., software subscriptions, maintenance contracts, gym]. Available data: [describe, e.g., subscription export with start date, end date, status, monthly amount]. Billing cycle: [monthly, annual, mixed].

First set the definition decisions, with the option you recommend for my case and its effect on the figure: churn in number of customers vs. revenue (both tell different stories: show a numerical example), the starting base (customers at start of period, including or excluding new ones), handling edge cases (pause, pending non-payment, downgrade, cancellation submitted but effective later, win-back customers).

Then provide the exact formula and its implementation in [Excel, Google Sheets, or SQL] on my data. Warn me against reading traps: mixing monthly and annual cycles in the same rate, churn that mechanically improves when sales accelerate, and why a tenure cohort reading usefully complements the global rate. Format: decisions, formula, implementation, traps.

I want to understand what each [product, service, or family] truly earns me. Data: [available export, e.g., sales by product with quantities and prices from POS, purchases and costs from accounting]. Identified costs: [list: purchases, materials, labor, fixed costs, shipping, commissions].

Build the analysis in stages: gross margin per product (selling price minus direct costs, specifying Net of Tax everywhere), then the tricky question of indirect costs: propose 2 plausible allocation keys for my activity (by revenue, by time spent, by volume), show how product ranking changes based on the key, and say why none is "the truth."

Report by crossing margin rate AND dollar margin with volume: the high-rate product that doesn't sell, the low-rate product that drives volume, and what we miss by looking at only one column. End with the 3 typical decisions this analysis can inform (pricing, discontinuation, promotion) and for each, the complementary figure to check before deciding. Format: method, sample table, commented reading.

I want to measure how much my business depends on a few [customers, products, channels, SKUs]. Data: [export, e.g., revenue by customer over 12 months from invoicing, sales by SKU from POS]. Tool: [Excel or Google Sheets].

Guide the construction of the concentration analysis: descending sort, cumulative total, and cumulative percentage (exact formulas in my tool), reading the meaningful thresholds (how many items for 50% and 80% of total, share of the top 1 and top 5), and a Pareto curve if useful (how to plot and read it).

Go beyond the finding: help me qualify the RISK of this concentration for my case: what happens if the top [customer or product] disappears (impact in dollars and percentage), is concentration increasing or decreasing compared to [available comparison period], and does the long tail (the many small ones) cost more to serve than it earns?

Constraints: handle related items that are actually one (same group, same buyer), and do not conclude "cut the tail" without looking at actual costs. Format: method, formulas, 5-line risk reading.

I want to compare the performance of my [locations, branches, teams, sites] without being unfair: they have different contexts. Units to compare: [list, with known differences: size, zone, tenure, headcount, foot traffic]. Data: [available metrics per unit and period].

Build the comparison in three steps. First, metrics NORMALIZED to a fair base (per salesperson, per square foot, per opening hour, per local resident: choose the relevant ratios for my activity and justify). Second, comparing each unit to itself (year-over-year growth): does the growth ranking tell a different story than the raw ranking? Finally, remaining inexplicable gaps: those that persist after normalization and deserve a visit or interview, not a penalty.

Constraints: flag units that are too new or too small to be ranked, never summarize in a single ranking without showing the component ratios, and list factors data doesn't capture (renovations, local competition, long absences). Format: sample comparative table, then reading grid for review with managers.

I want to analyze my conversion funnel to find where to focus effort. Journey: [describe steps in order, e.g., visit, product page, cart, shipping, payment, confirmation; or B2B: lead, contact made, meeting, quote, signature]. Volumes per step over [period]: [paste figures]. Measurement source: [GA4, CRM, other].

Calculate conversion rates between each consecutive step (not just end-to-end) and identify the step with the highest loss in absolute numbers AND in rate: both readings don't always point to the same step; explain which to prioritize for my goal [goal, e.g., more signatures with constant traffic].

Before concluding, check measurement traps with me: technical steps that don't always fire, people skipping steps or going backward, measurement window too short for slow journeys. Then propose the most instructive cross-tab with [available segment, e.g., mobile vs. desktop, traffic source, customer size] and what a gap between segments would imply. Format: rate table, diagnosis, 2 testable improvement hypotheses for the priority step.

I suspect a link between [variable 1, e.g., weather, number of online reviews, quote response time] and [variable 2, e.g., sales, signature rate] and I want to check it properly. Data: [paste the two paired series, with granularity: daily, weekly, per case]. Initial intuition: [what I believe and what I would do if confirmed].

Guide the analysis in [Excel or Google Sheets]: first the scatter plot and its reading (shape, direction, outliers pulling the relationship), then the correlation coefficient with its simple language interpretation and the thresholds beyond which it becomes noteworthy given my number of observations [number of points].

Then, the part that matters: review competing explanations before any causal conclusion: common factor influencing both variables (seasonality is prime for time series), reverse causality, pure coincidence on few points. For each, the simple test that strengthens or weakens it with my data. Mandatory three-line conclusion: what is established, what remains a hypothesis, and the concrete experiment that would settle it.